Scholarlink

<p>Background. High‑grade serous ovarian cancer (HGSOC) is a lethal disease marked by frequent platinum‑based chemotherapy resistance, resulting in high recurrence and mortality. Methods. Whole‑slide images (WSIs) and clinical data were retrieved from The Cancer Genome Atlas and our institutional database. After segmenting WSIs and discarding non‑informative tiles, a convolutional neural network (CNN) was trained to classify cancer versus normal tissue and to predict drug response. Clinical variables were integrated and feature selection performed using Lasso, AdaBoost, Naive Bayes, XGBoost and Random Forest, with the best‑performing model identified on validation and test sets. Resistance scores derived from the optimal model were correlated with clinicopathologic factors, and associations between individual features and lymphocyte infiltration were examined; key features were validated pathologically. Results. The CNN achieved an AUC of 0.995 for tumor‑normal discrimination and 0.662 for distinguishing resistant from sensitive tiles. All five machine‑learning models yielded area under the curves (AUCs) ≥ 0.90 for histologic features. Lasso performed best (AUC = 0.993) and selected 85 significant features. Higher resistance scores correlated positively with tumor grade and stage, negatively with silent mutation burden and neoantigen load, and positively with lymphocyte infiltration, especially feature TZ0279. Patients in the resistant group showed significantly poorer overall and disease‑free survival. Conclusion. A deep‑learning pipeline based on pathology images accurately separates tumor from normal tissue in HGSOC, and also has preliminary potential in predicting chemotherapy response. Histopathological features derived from Lasso regression can initially reflect the tumor microenvironment and drug resistance, offering potential biomarkers for prognosis and personalized therapy. This supports the feasibility of applying image-based artificial intelligence technologies to clinical decision-making research.</p><p>Clinical trial number. Not applicable.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Prediction of chemoresistance in ovarian cancer based on deep learning with pathological images

  • 期刊论文

  • 发表日期：

    2026年4月10日

查看全文
  • Leyan Niu,
  • Xiang Li,
  • Zhiquan Mao,
  • Fen Fu,
  • Mingjie Wang,
  • Hengbin Zhang,
  • Le Chen,
  • Qing Zhang,
  • Xiaoli Tang,
  • Weiming Lou

摘要

Background. High‑grade serous ovarian cancer (HGSOC) is a lethal disease marked by frequent platinum‑based chemotherapy resistance, resulting in high recurrence and mortality. Methods. Whole‑slide images (WSIs) and clinical data were retrieved from The Cancer Genome Atlas and our institutional database. After segmenting WSIs and discarding non‑informative tiles, a convolutional neural network (CNN) was trained to classify cancer versus normal tissue and to predict drug response. Clinical variables were integrated and feature selection performed using Lasso, AdaBoost, Naive Bayes, XGBoost and Random Forest, with the best‑performing model identified on validation and test sets. Resistance scores derived from the optimal model were correlated with clinicopathologic factors, and associations between individual features and lymphocyte infiltration were examined; key features were validated pathologically. Results. The CNN achieved an AUC of 0.995 for tumor‑normal discrimination and 0.662 for distinguishing resistant from sensitive tiles. All five machine‑learning models yielded area under the curves (AUCs) ≥ 0.90 for histologic features. Lasso performed best (AUC = 0.993) and selected 85 significant features. Higher resistance scores correlated positively with tumor grade and stage, negatively with silent mutation burden and neoantigen load, and positively with lymphocyte infiltration, especially feature TZ0279. Patients in the resistant group showed significantly poorer overall and disease‑free survival. Conclusion. A deep‑learning pipeline based on pathology images accurately separates tumor from normal tissue in HGSOC, and also has preliminary potential in predicting chemotherapy response. Histopathological features derived from Lasso regression can initially reflect the tumor microenvironment and drug resistance, offering potential biomarkers for prognosis and personalized therapy. This supports the feasibility of applying image-based artificial intelligence technologies to clinical decision-making research.

Clinical trial number. Not applicable.